huihui-ai/DeepSeek-V3-bf16

This model converted from DeepSeek-V3 to BF16.
Here we simply provide the conversion command and related information about ollama.

The following conversion also applies to deepseek-ai/DeepSeek-R1

If needed, we can upload the bf16 version.

FP8 to BF16

1. Download deepseek-ai/DeepSeek-V3 model, requires approximately 641GB of space.

cd /home/admin/models
huggingface-cli download deepseek-ai/DeepSeek-V3 --local-dir ./deepseek-ai/DeepSeek-V3

2. Create the environment.

conda create -yn DeepSeek-V3 python=3.12
conda activate DeepSeek-V3
pip install -r requirements.txt

3. Convert to BF16, requires an additional approximately 1.3 TB of space.

cd deepseek-ai/DeepSeek-V3/inference
python fp8_cast_bf16.py --input-fp8-hf-path /home/admin/models/deepseek-ai/DeepSeek-V3/ --output-bf16-hf-path /home/admin/models/deepseek-ai/DeepSeek-V3-bf16

BF16 to gguf

1. Use the llama.cpp (Download the latest version) conversion program to convert DeepSeek-V3-bf16 to gguf format, requires an additional approximately 1.3 TB of space.

python convert_hf_to_gguf.py /home/admin/models/deepseek-ai/DeepSeek-V3-bf16 --outfile /home/admin/models/deepseek-ai/DeepSeek-V3-bf16/ggml-model-f16.gguf --outtype f16

2. Use the llama.cpp quantitative program to quantitative model (llama-quantize needs to be compiled), other quant option. Convert first Q2_K, requires an additional approximately 227 GB of space.

llama-quantize /home/admin/models/deepseek-ai/DeepSeek-V3-bf16/ggml-model-f16.gguf  /home/admin/models/deepseek-ai/DeepSeek-V3-bf16/ggml-model-Q2_K.gguf Q2_K

3. Use llama-cli to test, llama-cli needs to be compiled.

llama-cli -m /home/admin/models/deepseek-ai/DeepSeek-V3-bf16/ggml-model-Q2_K.gguf -n 2048

Use with ollama

Note: this model requires Ollama 0.5.5

You can use huihui_ai/deepseek-v3:671b-q2_K directly

ollama run huihui_ai/deepseek-v3:671b-q2_K

or huihui_ai/deepseek-v3:671b-q3_K

ollama run huihui_ai/deepseek-v3:671b-q3_K